Predicting eutrophication dynamics using artificial neural networks

Abstract

Dal Lake, the freshwater lake in Srinagar, Jammu and Kashmir, has experienced significant water quality changes over the last two decades due to anthropogenic activities, intensifying eutrophication and threatening its ecological integrity. The study aimed to identify the main limiting factors of eutrophication process in Dal Lake and to rate the impact of traditional eutrophication drivers, i.e. total nitrogen (TN), ammonium (NH₄⁺), total phosphorus (TP), orthophosphate (PO₄³⁻), WT (water temperature), T (transparency), and chemical oxygen demand (COD), Artificial neural networks (ANNs) were employed to identify key drivers to develop predictive models of eutrophication dynamics. Three ANN models with different input combinations were developed. In an initial train–test split, the most complex configuration (model 3) yielded the highest correlation with the index of trophic state (ITS) (training R² = 0.24, testing R² = 0.20, r = 0.46–0.49), whereas 5-fold cross-validation showed that the simpler model 2 achieved the lowest average root mean square error (RMSE) and mean absolute error (MAE) but the highest mean R². Overall explanatory power was modest (R² ≤ 0.20), indicating that the ANNs captured only a small proportion of ITS variability. Permutation-based sensitivity analysis showed that COD and TP are consistently the most influential predictors of ITS, while the remaining nutrient variables (TN, PO₄³⁻, NH₄⁺) contributed weakly in this dataset. Thus, the ANN approach yields only partially informative insights into the relationships between water-quality variables and ITS in Dal Lake and is not yet suitable as a stand-alone forecasting tool for eutrophication dynamics.